A Two-Stream Graph Convolutional Neural Network for Dynamic Traffic Flow Forecasting

Forecasting the traffic flow is a critical issue for researchers and practitioners in the field of transportation. Using the graph convolutional network (GCN) is widespread in traffic flow forecasting. Existing GCN-based methods mostly rely on undirected spatial correlations to represent the features of spatial-temporal graph. What's more, the traffic flow renders two types of spatial correlations, including the stable correlation constrained by the fixed road structure and the dynamic correlation influenced by the traffic fluctuation. In this paper, we propose a two-stream graph convolutional network by considering stable and dynamic correlations in parallel, which is an end-to-end deep learning framework for dynamic traffic forecasting. We present an auto-decomposing layer to decompose real-time traffic flow data into a stable component and a dynamic component with different spatial correlations. Specifically, the stable component is constrained by the physical road network, and the dynamic component represents fluctuations caused by changes in traffic conditions such as congestion and bad weather. Moreover, we extract stable and dynamic spatial correlations through our two-stream graph convolutional layer. Finally, we use parameterized skip connection to fuse spatial-temporal correlations as the input of output layer for forecasting. Extensive experiments are conducted on two real-world traffic datasets, and experimental results show that our proposed model is better than several popular baselines.